DC-7: Mohammadreza Jahanbakhsh

Contact

E-mail: mjahanbakhsh24@ku.edu.tr

Project

Physics-aware machine learning for control of chemical processes

Host organization

KU

Supervisors

Prof. Can Erkey (Main, KU); Prof. Gürkan Sin (co-supervisor, DTU)

Duration

36 months

Objectives

Physics-informed knowledge will be integrated into the offline and online training instances of the neural networks to improve test performance and reduce the possibility of suboptimal training. Those algorithms will be developed for designing the most proper architecture of machine (deep) learning algorithms (e.g., recurrent neural network (RNN) or long short-term memory (LSTM) for dealing with time series data, among others).

A systematic framework to build, update, and validate novel predictive analytics algorithms using large time series data sets from production plants will be developed. Trained and validated machine learning models will be embedded into model-based closed-loop control algorithms such as classical, nonlinear, and economic model predictive control over formulating tailored optimal control problems. The extension of physics-informed machine learning methods for reinforcement-learning-based control will also be investigated. The results will be applied to a case study from urea including the hydrogen processing industry.

Expected results

  1. Physics-informed machine learning algorithms to control chemical processes: Economical benefits will also be included in control objectives in a multi-objective manner to find the best trade-off between safety and economic optimality for industrial plants such as urea and gas and hydrogen-based production,
  2. Physics-informed (reinforcement) learning algorithm capable of learning the best mitigation actions for safety and economic optimality,
  3. Closed-loop performances of standard-modular, first principles-based and machine learning-based control structures will be compared based on safety and economics,
  4. Application of the deployed algorithms for real-time operation on urea including hydrogen processing industries.

Planned secondments

  1. DTU, M13, 2 months: Training on advanced uncertainty and sensitivity analysis
  2. NOVOTEC, M30, 2 months: Training on case study training and application for LNG operations